LEARNING BIAS IN NEURAL NETWORKS AND AN APPROACH TO CONTROLLING ITS EFFECTS IN MONOTONIC CLASSIFICATION

As a learning machine, a neural network using the back-propagation training algorithm is subject to learning bias. This results in unpredictability of boundary generation behavior in pattern recognition applications, especially in the case of small training sample size. This research sugests that in a large class of pattern recognition problems, such as managerial and other problems possessing monotonicity properties, the effect of learning bias can be controlled by using multiarchitecture monotonic function neural networks.